Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 36: Condensed-matter simulations augmented by advanced statistical methodologies (joint session DY/CPP)
DY 36.4: Talk
Wednesday, April 3, 2019, 16:00–16:15, H20
Representing molecules and materials for accurate interpolation of quantum-mechanical calculations — •Marcel Langer, Alex Goessmann, and Matthias Rupp — Fritz Haber Institute of the Max Planck Society, Berlin, Germany
The search for novel materials, the exploration of phase diagrams and other high-throughput applications require numerical simulations of molecules and materials from first principles, but are limited by their high computational cost. By interpolating between reference calculations, machine learning can act as a fast accurate surrogate for these calculations, greatly increasing the number of accessible systems. [1] This requires a representation of molecules and materials suitable for interpolation. We show how the state-of-the-art representations can be understood in a unified framework [2] based on local descriptions of atomic environments via k-body functions, group averaging and tensor products, and discuss implications. We benchmark predictive accuracy of selected representations by carefully controlling for all other factors, including data distribution, regression method and optimization of free parameters. For the latter, we employ a consistent and fully automatic procedure to optimize both numerical and categorical free parameters, such as the choice of k-body functions, using sequential model-based optimization with tree-structured Parzen estimators. [3]
References: [1] a) Rupp et al, Phys Rev Lett 108, 058301, 2012. b) Rupp, Int J Quant Chem 115, 1058, 2015. [2] Willatt et al, Phys Chem Chem Phys, accepted, 2018. [3] a) Bergstra et al, NIPS 24, 2546, 2011; b) Bergstra etal, ICML 30, I-115, 2013.